The main purposes of document analysis, which include error correction, semantic similarity assessment, and duplication identification, are essential to maintaining the authenticity and uniqueness of textual material. The hybrid deep learning method presented in this research computes fine-grained semantic similarity between phrase pairs by combining the T5 Transformer and BiLSTM. The model generates continuous similarity indices that may detect minute differences in tampered or paraphrased text after being trained using Mean Squared Error (MSE) loss on an improved SQuAD v1.1 dataset featuring 18,891 sentence pairs labeled with similarity scores. With a Spearman correlation of 0.8769, Pearson correlation of 0.9424, MSE of 0.0254, and an F1-score of 0.9028, the suggested T5–BiLSTM architecture demonstrated its efficacy in document similarity tasks. These outcomes demonstrate the method’s usefulness in applications including academic grading, plagiarism detection, and semantic document comparison.